DeepSeek V4.1 Flash Launches on China's National Supercomputing Internet
DeepSeek V4.1 Flash, a 552-billion parameter Mixture-of-Experts (MoE) model, has officially launched on China's National Supercomputing Internet platform via API access. Featuring a novel asymmetric Causal-Encoder-Decoder architecture, the model outperforms previous flagships like DeepSeek V4 Pro on benchmark intelligence tests. By dramatically compressing KV Cache requirements—reducing HBM usage by 75% and SSD memory demand by 87.5%—the model significantly slashes inference costs for memory-intensive AI agent workflows. This architectural leap offers a far more cost-effective solution for deploying massive LLMs in production environments. The architecture utilizes asymmetric input and output activations, employing only 8B active parameters for inputs and 16B for outputs to keep operational costs low. It achieves enhanced capability through novel pre-training techniques combined with large-scale reinforcement learning (RL) post-training.
## BACKGROUND
Mixture-of-Experts (MoE) models route inputs to specialized sub-networks, allowing high total capacity while maintaining lower computational costs per request. Meanwhile, the KV Cache stores previous attention tokens during text generation to avoid redundant computation, but it consumes vast amounts of high-bandwidth GPU memory (HBM), creating significant bottlenecks for long-context applications.